Skip to content

Release 1.0.0: pre-tag audit fixes and precise licences (hold until SDE passes) - #111

Merged
Pixel-Minions merged 13 commits into
mainfrom
card/release-1.0.0-fix
Nov 5, 2026
Merged

Pixel-Minions merged 13 commits into
mainfrom
card/release-1.0.0-fix

Conversation

@Pixel-Minions

@Pixel-Minions Pixel-Minions commented Nov 3, 2026 •

Copy link
Copy Markdown
Owner

Pre-tag audit fixes on top of the release commit ec1f768d. The full SDE run on ec1f768d (run 41658641028, two attempts) plus local SDE runs of what it did not complete qualify every CPU-dependent test on all five emulated CPUs, on both platforms (recorded in docs/verification.md). The merge commit is the commit to tag.

Nothing here changes what the port computes: doc comments, docs, package metadata, xtask's wheel check, re-exports and a test split, so the SDE results of ec1f768d carry over (release-checklist step 5: no CPU profile is affected).

  • CHANGELOG 1.0.0 lists the accepted speed misses; known-issue sentences say which crash each describes.
  • Release checklist: wheel upload step.
  • I-158 retired; W0001/W0002 sections in docs/waivers.md.
  • README status and M5 row for 1.0.0; CPython in the licence list.
  • -ops, -formats, -gpu say on docs.rs that their APIs are internal.
  • Rustdoc: 40 unclosed-HTML and 20 unresolved-link warnings fixed (doc comments only).
  • Licences (owner decision 2026-11-03): ops BSD-3-Clause AND Apache-2.0 WITH LLVM-exception, formats BSD-3-Clause AND MIT, gpu BSD-3-Clause, opencolorio-rs BSD-3-Clause AND MIT AND Zlib (SampleICC's ICC licence, which has no SPDX id, named in its README); the wheel's PEP 639 expression covers everything it holds and links, with LicenseRef-ICC-0.2.

Checks (Windows): gate --crates ops/formats/gpu/opencolorio-rs/xtask (2989 passed), guards, ci, cargo-deny licenses, cargo package of the four crates, cargo xtask wheel-check (384/384, tools match).

  • CpuProcessor and its image types named in ocio (5552e2c): CpuProcessor, the image descriptions and the dynamic property types are re-exported at the crate root, so a user doesn't need opencolorio-rs-ops to name them. Tested by tests/public_names.rs and a doctest. Re-exports only: the SDE results still apply.

  • Slow legacy GPU sweeps split (2f31255): P33 into four parts and P34 into three, each with an audit that every original request runs once, in order; the SDE schedule maps the parts. Test grouping only. All parts passed under local SDE (P34 on nhm, skl, skx; P33 on skl).

  • README cleanup (cae4c4a): 360 lines down to 115; the verification detail moves to docs/verification.md, the phase-by-phase history to docs/history.md.

  • SDE qualification recorded (59abb91) in docs/verification.md.

🤖 Generated with Claude Code

Pixel-Minions and others added 13 commits November 3, 2026 13:27
CHANGELOG.md, 1.0.0: list the speed misses accepted for 1.0.0 from
docs/cards/phase8.md:115-128 (apply_rgba for Log and GradingPrimary, ff_red_mod_10,
lut1d u10 to f32, and the Linux-only misses) and say what the ratios mean, as release
checklist step 6 requires. Repository paths become full GitHub URLs, since the note
becomes the release page. Known issues: the sentences about the native replay and the
refused boundary requests now say they are about the snb crash; no fact changes.

Checks: cargo xtask guards, cargo xtask ci (Windows).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
docs/release-checklist.md: step 8, after the crates: the owner uploads the Windows
cp313-cp313-win_amd64 and Linux cp313-cp313-manylinux_2_34_x86_64 wheels from the tagged
commit's CI run artifacts (kept 30 days) to PyPI as opencolorio-rs; the first upload needs
an account-wide token, since a project-scoped one exists only after the project does.

Checks: cargo xtask guards, cargo xtask ci (Windows).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
docs/improvements.md, I-158: the archive reader landed in p4-baker (4.9a1, 4ddda22), and
the "not ported yet" error is gone from crates/ocio/src. The entry is marked retired,
naming that card, the ported OCIOZArchive test that reads archives and
crates/ocio/tests/ocioz_read_oracle.rs. Its number stays.

Checks: cargo xtask guards (register structure), cargo xtask ci (Windows).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
docs/waivers.md: short sections for W0001 (PQ on the exact-math path, Windows) and W0002
(NaN transform parameters), from waivers.toml's own text and CLAUDE.md's "W0002 only
through the battery"; the intro no longer speaks of "this proposal" and says that
waivers.toml holds where the two differ. waivers.toml is unchanged.

Checks: cargo xtask guards, cargo xtask ci (Windows).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
README.md: the status block and the M5 row say that 1.0.0 (1.0.0+ocio.2.5.2) is the
release, on crates.io as opencolorio-rs (with -ops, -formats, -gpu) and on PyPI as
opencolorio-rs, and that M5 is done; "Nothing is published yet" (three places) and
"Release-audit findings and packaging also remain open" are gone. The licence list names
CPython, as THIRD-PARTY-NOTICES.md does.

Checks: cargo xtask guards, cargo xtask ci (Windows).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…re internal

crates/ocio-ops, ocio-formats, ocio-gpu: a crate-level //! paragraph and the same sentence
in each package README: these crates are building blocks of opencolorio-rs, use that
crate, and their APIs follow OpenColorIO's internals and may change in minor releases.
The only statement before was a plain // comment in ocio-ops/src/lib.rs, which rustdoc
doesn't show. Doc comments and READMEs only.

Checks: cargo xtask guards, cargo xtask ci, cargo fmt --all --check, cargo clippy
--workspace --all-targets -- -D warnings, cargo test --doc -p opencolorio-rs (Windows).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Doc comments only (/// and //!), in opencolorio-rs and opencolorio-rs-ops: placeholders
in quoted upstream messages such as '<filepath>' or <index> are escaped (\<name>), so
they render instead of vanishing as unclosed HTML tags; bracketed values such as [0,1]
and [istream] are escaped, LUT[start] expressions go in backticks, and the links to
KnotsCoefs::eval_curve(_rev) and CtfReaderTransformPtr get their paths. The bare
yaml.org URL in yaml_cpp/convert.rs becomes an autolink. No code, string literal or
message text changes.

cargo doc --no-deps -p <crate>, before -> after (unclosed HTML, unresolved links):
opencolorio-rs 32, 2 -> 0, 0; opencolorio-rs-ops 8, 18 -> 0, 0; -formats and -gpu
0, 0. Links to private items (15, 44, 2, 1) and one redundant explicit link target in
ops are left as they were.

Checks: cargo xtask guards, cargo xtask ci, cargo fmt --all --check, cargo clippy
--workspace --all-targets -- -D warnings, cargo test --doc -p opencolorio-rs (Windows).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Owner decision 2026-11-03 (pre-tag audit, SHOULD 1): each published
package's license field names the licenses of the code it holds, not only
BSD-3-Clause. crates.io takes only SPDX-listed licenses, so SampleICC's ICC
Software License 0.2 (no SPDX identifier) is named in words in the
opencolorio-rs README (owner's choice over `license-file`).

- opencolorio-rs-ops: BSD-3-Clause AND Apache-2.0 WITH LLVM-exception
  (the MSVC STL <regex> port; Imath and pystring are BSD-3-Clause).
- opencolorio-rs-formats: BSD-3-Clause AND MIT (expat).
- opencolorio-rs-gpu: BSD-3-Clause (unchanged, the workspace's).
- opencolorio-rs: BSD-3-Clause AND MIT AND Zlib (yaml-cpp, minizip-ng),
  SampleICC named in its README.
- The wheel (PEP 639, which allows LicenseRef-): BSD-3-Clause AND
  Apache-2.0 WITH LLVM-exception AND MIT AND Zlib AND BSL-1.0 AND PSF-2.0
  AND Unicode-3.0 AND LicenseRef-ICC-0.2 AND (MIT OR Apache-2.0) AND
  (BSD-2-Clause OR Apache-2.0 OR MIT): the port, what it translates
  (CPython's PSF-2.0 in the module, Little CMS's MIT in the tools), and the
  licenses of the crates `cargo metadata` shows it links.
- xtask wheel-check holds METADATA to the new expression
  (`LICENSE_EXPRESSION`), and a new test holds that to pyproject.toml.
- The package READMEs and the root README's License section say the same.
The private crates keep the workspace's BSD-3-Clause; nothing publishes them.

Checks (Windows): cargo xtask guards; cargo xtask ci; cargo test -p xtask
(146 passed); cargo deny check licenses (ok); cargo package --offline of the
four crates (packaged and verified; the packaged Cargo.toml files carry the
new fields); cargo xtask wheel-check (wheel 1.0.0 built, License-Expression
as above, 384 of 384 Python tests and the tools match the wheel).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
`opencolorio-rs` handed out `CpuProcessor` (`Processor::default_cpu_processor` and its
siblings) without re-exporting it, and `CpuProcessor::apply` takes image descriptions that
weren't re-exported either, so a user needed the internal `opencolorio-rs-ops` crate to name
them. OCIO declares all of them in its one namespace (include/OpenColorIO/OpenColorIO.h @
v2.5.2), as the crate already does for `GpuProcessor`.

Re-exported at the crate root:
- `CpuProcessor`;
- the image descriptions: `PackedImageDesc`, `PlanarImageDesc`, the `ImageDesc` and
  `ImageDescMut` traits, `ChannelOrdering`, and their helpers `PixelData`, `PixelParts`,
  `Bytes`, `At`, `ImageLayout` and `ChannelPos`;
- the dynamic properties: `DynamicPropertyRcPtr`, `DynamicPropertyImpl`, each kind's `Impl` and
  `ImplRcPtr`, and the grading states their `state()` returns.

A scan of every public function signature in the crate finds no other type from
`opencolorio-rs-ops` it doesn't name, except the CTF reader's internals, which aren't OCIO API.

Tests: `tests/public_names.rs` names each re-export through `ocio::` only, and renders an image
through a dynamic exposure with them; the crate docs gain a whole-image example (a doctest).
CHANGELOG says `opencolorio-rs` alone is enough.

Evidence: `cargo xtask gate` on Windows: fmt, clippy, ci, 3557 tests, 216 Python tests and the
Python suite pass. No compiled code changes, so the SDE results on ec1f768 still apply.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
… parts

The full SDE run on ec1f768 (41658641028) found no mismatch, but on Windows the whole P34
sweep (`legacy_dynamic_positions_and_snapshot_values_match_native_wheel`) took the per-test two
hours on nhm, skl and skx, and the whole P33 sweep
(`legacy_allocation_cuts_and_file_boundaries_match_native_wheel`) on skl; the timeouts then
used up those jobs' budgets. Owner, 2026-11-04: shard it.

As P32 and P35 already are, each plan is generated complete, then run in contiguous parts in its
original order: P33 (980 requests) in four parts of 245, P34 (1680) in three of 560. Every
comparison is unchanged, and each P34 part still requires dynamic snapshot rows.
`allocation_legacy_parts_cover_original_requests_once` and
`dynamic_legacy_parts_cover_original_requests_once` check that the parts are the original plan,
each input once and in order. In the 16-shard SDE matrix the new names land in shards 1, 8 and 2
(P34) and 9, 9, 4 and 14 (P33), instead of 11 and 5. `scripts/sde_cpu_tests.py`'s four-way map
and `p50_schedule_controls.py` name the parts.

Evidence: all 11 split and audit tests pass natively against the wheel (Windows, release);
`p50_schedule_controls.py` passes; `cargo xtask gate --crates opencolorio-rs` on Windows
passes (1835 tests).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The README had grown into a status log. It now says what the project is, how to install and
use it, what is included, how exact it is and how to build it, in 115 lines instead of 360.
The detail moves to two new pages, with nothing dropped:

- docs/verification.md: the parity rules, the oracle, the battery, the platforms, the
  emulated CPUs, the upstream test accounting and the waivers (W0001 and W0002 in full).
- docs/history.md: what landed in each milestone and phase, M0 to M5.

Stale status lines ("next:", "remains open") are gone; the changelog keeps the speed
exceptions and the known issues.

Evidence: cargo xtask ci (guards, fixtures, ratchet) passes; every relative link in the
three files resolves. Docs only, so no test step.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
docs/verification.md gains the 1.0.0 qualification: SDE run 41658641028 on ec1f768 (two
attempts, every Rocky job passed, every Windows failure a job-budget timeout or the known
wheel crash under emulation), and the local SDE runs on 2f31255 that cover what neither
attempt completed: the P33 and P34 split parts, and 26 tests on hsw, skl, nhm and skx.

Evidence: the per-test reports of both CI attempts, combined per CPU and shard against the
complete test list; the local runs passed against the official wheel with
OCIO_RS_ORACLE_NO_CACHE=1. cargo xtask guards passes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…eps, short README, SDE qualification recorded

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@Pixel-Minions
Pixel-Minions merged commit 44b5db9 into main Nov 5, 2026
43 checks passed
@Pixel-Minions
Pixel-Minions deleted the card/release-1.0.0-fix branch November 5, 2026 05:47
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant